AI Generation

AI generation lives in two places, and which one you are in decides whether you get to review the result first.

The five generators — the forms that make something — are tabs in Thalia’s dock, down the right edge of the workspace (the ✦ Thalia button in the header, or Ctrl+I). Its strip reads Chat · Dialogue board · Definition · Entities · Picture · Variables │ History; the icons carry their names as tooltips, and the empty chat lists the five in words. One form is on screen at a time, and its result lands as a card in the chat thread, where you review it before anything is written.

What is not a form is a page that opens as a tab beside your boards: three read-only meters, listed under AI Generation on the rail, and one batch job, AI batch translation, listed under Localization with the rest of the translation work.

Thalia herself — the Chat tab of that same dock — is a different surface with different behaviour, and the difference matters: she writes to your project as the agent finishes. See Thalia, the assistant dock.

The five generators, in the dock

Tab	What it makes	What you give it
Dialogue board	A branching node graph	A prompt, a mode (free or from a document), @-mentions, a retrieval toggle, a node-type scope, an orientation
Definition	A typed entity schema — fields, types, defaults	A prompt, @-mentions, a tools toggle, a model
Entities	A populated entity item against a definition you choose	A definition, a prompt, an image checkbox, a retrieval toggle, @-mentions
Picture	Concept art, portraits, scene imagery	A prompt, a model and preset, an aspect ratio or explicit size
Variables	Several game variables at once	One definition per line, as name: type = default

The four pages

Each is a row of a list on the left; clicking the row opens the page as a tab.

Page	Listed under	What it does	What you give it
AI batch translation	Localization	Translations and keyed string resources	Two panels; a model tier or a custom model
Embeddings	AI Generation	Manages the retrieval index	—
Usage & Costs	AI Generation	A read-only dashboard	—
Model Credits	AI Generation	A read-only price table	—

Two more generations exist that have no form or page: an AI cell in a spreadsheet, invoked from the spreadsheet itself, and a board polish pass.

What the AI can see

For a board or entity generation, what is assembled and sent is:

Always sent
Your @-mentioned entities	In full
Your @-mentioned documents	The first 5,000 characters of each, from the stored text; anything longer is cut there, and the model is told it was
The project’s variables and variable contexts
The available boards
The available commands
A count of media	The list itself sits behind a tool the model must ask for

Sent only when retrieval is on
A retrieved lore block	Found by similarity against your project

An entity generation receives only the first two rows — your mentioned entities and documents — and the retrieved block when retrieval is on.

In document mode the source document’s text is the primary block, sent whole rather than cut at 5,000 characters, and the entity list collapses to names — a hint for a tool that resolves characters on demand, rather than the full records.

So the honest sentence is: retrieval is opt-in; what you explicitly attach and the project’s own rosters are always sent.

Review before commit

Every generator stages. Its result appears as a card in the chat thread with Apply, Discard and Regenerate, and nothing reaches your project until you press Apply.

Tab	What Apply writes
Definition	The entity definition
Entities	The entity item
Dialogue board	The board, its nodes and its edges
Variables	The variables
Picture	The image, into your media library

Which models

The model is chosen per generation, from a picker that covers several sizes of general-purpose model. The list is served live rather than baked into a release, so what the picker shows is the current one — this page deliberately does not reproduce it.

Images use a separate lineup of image models, each with its own presets and aspect-ratio options.

The practical guidance is short: a larger model is not automatically a better one here. Most of these generations are structured — a schema, a node graph, a set of variables — and a mid-size model usually produces the same structure for a fraction of the credits. Try the cheaper one first and move up only when the output is actually wrong.

Credits

Every generation spends credits. Your plan includes an allowance each month, and you can buy more in packs.

Every generation records the model it used and the credits it spent, so nothing is a mystery after the fact. A generation that never reaches a model — the Variables tab, for instance — costs nothing at all.

How the balance is spent

Monthly credits are spent first, then any you have purchased.

• Unused monthly credits do not roll over. Purchased credits do persist.
• The monthly allowance resets on your own billing cycle.
• A member with the Viewer role is allocated none.

If a generation starts while you have very little left, it still finishes and is delivered — your balance simply lands at zero rather than the work being thrown away after the model had already run.

Your allowance and the credit packs are on the pricing page, which is the one place plan figures are published.

Limits

Limit	Value
Rate	10 requests per 60 seconds, per IP address — not per user
Agent loop	12 iterations
Credits	Per plan; see pricing

AI capability itself is not plan-gated, and no individual generation type is either. Credits are the meter; every plan including Free can use every generator and every AI page.

What breaks

Situation	Message
Out of credits, before the call	Insufficient AI credits. Please purchase more credits or wait for your monthly reset.
Out of credits, at the ledger	Insufficient AI credits
Image generation is unavailable	A message saying image generation is not configured — contact support
A generation service is unavailable	A message saying AI generation is not configured — contact support
Every cell in a batch failed	AI batch cell generation failed: «first error» — and no charge is taken
Too many requests	Rate limited at 10 per minute per IP

A generation you have already received is never taken away by a billing problem.